Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution (CLIT)

This repository contains the PyTorch based official implementation of the paper titled:
Cascaded Local Implicit Transformer for Arbitrary-Scale Super-Resolution CVPR 2023.

Dependencies

• Python >= 3.7.0
• PyTorch >= 1.8.0

Train

EDSR-Baseline

Stage1: python train.py --config configs/train/train_edsr_baseline_lit.yaml --name lit_edsr

Stage2: python train.py --config configs/train/train_edsr_baseline_clit2.yaml --name clit_edsr2

Stage3: python train.py --config configs/train/train_edsr_baseline_clit3.yaml --name clit_edsr3

RDN

Stage1: python train.py --config configs/train/train_rdn_lit.yaml --name lit_rdn

Stage1: python train.py --config configs/train/train_rdn_clit2.yaml --name clit_rdn2

Stage1: python train.py --config configs/train/train_rdn_clit3.yaml --name clit_rdn3

SwinIR

Stage1: python train.py --config configs/train/train_swinir_lit.yaml --name lit_swinir

Stage2: python train.py --config configs/train/train_swinir_clit2.yaml --name clit_swinir2

Stage3: python train.py --config configs/train/train_swinir_clit3.yaml --name clit_swinir3

$\ast$ Please note that, if you want to cascadedly train stage2 or stage3 CLIT, you need to modified the "pre_train" property in the configuration so as to load previous stage1 or stage2 model as the pre-trained model.

Ex: train the stage2 CLIT using edsr-baseline model

pre_train: save/lit_edsr/epoch-last.pth

Test

EDSR-Baseline or RDN

bash eval.sh "put the model name here"

SwinIR

bash eval_swinir.sh "put the model name here"

Demo Attention Maps

python demo.py --model save/lit_rdn/epoch-last.pth --img_path assests/0868x4.png --scale 6

Inputs	Attention Heads

Additional Quantitative Results

Div2k

Method (SSIM)	x2	x3	x4	x6	x12	x18	x24	x30
EDSR-Baseline-CLIT	0.9397	0.8790	0.8266	0.7503	0.6439	0.6006	0.5771	0.5629
RDN-CLIT	0.9418	0.8829	0.8319	0.7564	0.6497	0.6053	0.5804	0.5657
SwinIR-CLIT	0.9436	0.8859	0.8357	0.7608	0.6534	0.6080	0.5830	0.5675

Set5

Method (SSIM)	x2	x3	x4	x6	x8
RDN-CLIT	0.9474	0.9101	0.8760	0.8053	0.7451
SwinIR-CLIT	0.9482	0.9117	0.8787	0.8131	0.7521

Set14

Method (SSIM)	x2	x3	x4	x6	x8
RDN-CLIT	0.9023	0.8227	0.7619	0.6748	0.6184
SwinIR-CLIT	0.9030	0.8262	0.7656	0.6789	0.6210

B100

Method (SSIM)	x2	x3	x4	x6	x8
RDN-CLIT	0.8962	0.8003	0.7304	0.6404	0.5876
SwinIR-CLIT	0.8975	0.8029	0.7341	0.6443	0.5907

Urban100

Method (SSIM)	x2	x3	x4	x6	x8
RDN-CLIT	0.9298	0.8568	0.7942	0.6918	0.6224
SwinIR-CLIT	0.9335	0.8651	0.8051	0.7070	0.6369

Acknowledgements

This repo is built on LIIF and LTE. Thanks the authors for their contributions and generosity.